podman load is very slow on macos

[contagnas]

On macos podman load seems much slower than it should be. sshing to the podman machine and running the load from there is much faster. For example, podman load on a tensorflow image took 4.75 minutes while podman machine ssh -- podman load took 30 seconds on my machine.

Reproduction steps:

podman pull tensorflow/tensorflow:latest-gpu
podman save tensorflow/tensorflow:latest-gpu -o tensorflow.tar
podman image rm tensorflow/tensorflow:latest-gpu
time podman load -i tensorflow.tar                                       # takes 4 minutes
podman image rm tensorflow/tensorflow:latest-gpu
time podman machine ssh -- podman load -i $(realpath tensorflow.tar)     # takes 30 seconds

Environment (rootful vs rootless doesn't seem to make a difference):

~ $ podman machine info
host:
    arch: arm64
    currentmachine: podman-machine-default
    defaultmachine: ""
    eventsdir: /var/folders/tr/bx7bxwwd1fx373hqgjrz4d980000gn/T/storage-run-501/podman
    machineconfigdir: /Users/chills/.config/containers/podman/machine/applehv
    machineimagedir: /Users/chills/.local/share/containers/podman/machine/applehv
    machinestate: Running
    numberofmachines: 1
    os: darwin
    vmtype: applehv
version:
    apiversion: 5.4.1
    version: 5.4.1
    goversion: go1.24.1
    gitcommit: b79bc8afe796cba51dd906270a7e1056ccdfcf9e
    builttime: Tue Mar 11 11:37:40 2025
    built: 1741718260
    buildorigin: pkginstaller
    osarch: darwin/arm64
    os: darwin

~ $ podman machine inspect
[
     {
          "ConfigDir": {
               "Path": "/Users/chills/.config/containers/podman/machine/applehv"
          },
          "ConnectionInfo": {
               "PodmanSocket": {
                    "Path": "/var/folders/tr/bx7bxwwd1fx373hqgjrz4d980000gn/T/podman/podman-machine-default-api.sock"
               },
               "PodmanPipe": null
          },
          "Created": "2025-05-14T15:39:12.161613-07:00",
          "LastUp": "2025-05-15T11:05:04.838064-07:00",
          "Name": "podman-machine-default",
          "Resources": {
               "CPUs": 5,
               "DiskSize": 100,
               "Memory": 2048,
               "USBs": []
          },
          "SSHConfig": {
               "IdentityPath": "/Users/chills/.local/share/containers/podman/machine/machine",
               "Port": 52429,
               "RemoteUsername": "core"
          },
          "State": "running",
          "UserModeNetworking": true,
          "Rootful": true,
          "Rosetta": true
     }
]

[baude]

how large is the image you are loading. when issuing from the mac, it needs to transfer the image first into the VM.

[mheon]

I suspect that is the cause - we're using the remote client (so transferring into the VM) even if it's a file available on the VM via mount.

[contagnas]

how large is the image you are loading. when issuing from the mac, it needs to transfer the image first into the VM.

in that example the tensorflow is something like 7GB (I picked the largest image I could find)

I suspect that is the cause - we're using the remote client (so transferring into the VM) even if it's a file available on the VM via mount.

This was my guess, which is how I arrived at this podman machine ssh -- podman load trick. This is working for me but it'd be nice if the podman cli could do this transparently.

[mheon]

If we were to develop such a technique, it could be used for build as well - probably see some pretty significant performance improvements with it there.

[Honny1]

This looks like good material for my thesis. Can I have it?

[mheon]

If you want it, feel free. Could be a big PR though.

[github-actions]

A friendly reminder that this issue had no activity for 30 days.